Source: Two AI Search weekly news episodes, YouTube auto-captions fetched 2026-10-30: raw/Quantum_breakthrough_Gemini_4_Argon_MOSS_Image_KLING_4_OpenAI_spy_Muse_Wonder_-_AI_NEWS.md (https://www.youtube.com/watch?v=tDWtcyeNNew, uploaded 2026-10-18, about 41 min) and raw/Cybersecurity_chaos_AI_for_DNA_Kimi_K3.1_Claude_Duet_Dessa_Fibo_2_Minimax_3.1_-_AI_NEWS.md (https://www.youtube.com/watch?v=EfyQOAJjx1A, uploaded 2026-10-25, about 43 min). Both sponsored by Luma. 2026-11-01 update adds Matt Wolfe’s episode for the same week: raw/AI_News_-_24_Things_You_Need_To_Know_This_Week.md (https://www.youtube.com/watch?v=TVRCcMlvR10). Item names are spelled as on the creator’s link list; caption errors such as “GauGAN animate” (GaussiAnimate), “Uniform 3D” (UniForge 3D), “Kimi Chat 3.1” (Kimi K3.1) and “Crea / Creia” (Krea) are corrected.
Two weeks of AI Search’s news roundups, filtered to what a marketer or builder can use: open models you can download today (TTS, image editing, transcription, 3D, world models), hosted releases (Gemini 4 Argon, Meta’s Muse Wonder, Kimi K3.1 Pro, Kling 4, Krea 3, Ideogram Vector, Claude Duet), and a wave of security incidents that touched tools developers use, including Railway and LiteLLM. Benchmark claims are the vendors’ or third-party leaderboards read on screen; the presenter’s opinions are marked. Pure research and robotics demos are listed briefly at the end.
Key Takeaways
- Rotate secrets if you use Railway, LiteLLM or OpenCode. AI Search reports attackers decrypted some Railway customer environment variables (Railway paused builds and deploys), pushed a malicious LiteLLM release from three developers’ hijacked GitHub logins (pulled within about 2–3 hours), and exposed API keys for 1.4% of OpenCode users. Also patch PHP, GitLab (a 10/10 severity flaw) and Node.js servers.
- New open TTS leaders: Fish Audio’s Drama 3 (#1 open model on Artificial Analysis, 35B total / 3B active, reported to run on 8 GB, emotion meta tags, voice cloning) and Ant Group’s Ming 3.0 TTS (MIT licence, 3 GB and 18 GB versions, the larger one does multi-speaker podcasts, slightly above Drama 3).
- New open image models: MOSS Image (5B, Apache 2.0, commercial use allowed, layout boxes; tutorial in Local Image Models), Fibo 2.0 by Bria (#1 open image editor on Artificial Analysis, just above Qwen Image 2.1, but non-commercial), and FLUX 3 Image Klein (5.6B, 4K, bounding boxes, but “anatomy horror” in his tests and a non-commercial community licence).
- Tiny, fast transcription: Bark3 (163 MB small, up to 1.25 GB Pro) claims Whisper Large v3 / Parakeet-level accuracy at a fraction of the size, commercial use with attribution.
- Hosted creative tools: Kling 4 (10 keyframes, 15 references, stereo audio, 10-bit HDR, 4K coming; he still ranks Seedance higher), Krea 3 (2K edits in under 8 s against 30 s+ for Nano Banana, up to six references, may be open-sourced), Ideogram Vector (prompt or image to editable vector graphics).
- Frontier and open LLMs: Gemini 4 Argon (1M output tokens, #1 on Vals and BlueprintBench, Pro/Ultra or paid API only), Meta Muse Wonder (fast, honest on hallucination tests, expensive per task) and Muse Spark Open (816B open weights with a restrictive licence), Kimi K3.1 Pro (about 3T parameters, above Fable 5.1 on the AA index at about $2 per task, open weights promised November 4), MiniMax M3.1 Flash, Step 5 Preview, Nemotron 3.5 Super VL (600+ tokens/s), LFM2.5-24B-A2B.
- Claude Duet is Anthropic’s proactive assistant on Opus 5.5 with its own cloud computer and 6,000+ integrations; waitlist, US, Pro or Max. Full coverage in Claude Duet.
Security incidents, week of October 25 (as AI Search reports them)
Not independently checked here; confirm with each vendor’s advisory.
- Framing: researchers used frontier models to find long-hidden bugs while attackers use AI to automate exploitation. He shows a chart claiming “roughly 3% of all critical software bugs in history have been found in just the last month.”
- PHP: two AI-found vulnerabilities allowing unauthenticated remote code execution; some sites, including government sites, went offline as a precaution. Patches are out.
- GitLab: severity 10/10 unauthenticated code execution; no confirmed widespread attacks; patched.
- Node.js: multiple vulnerabilities, patched.
- eMudhra (certificate authority, widely used in India): signing keys were “ridiculously weak” (one was the number four), so certificates could be forged. Keys revoked; eMudhra says it will leave the certificate business.
- Railway: attackers reached internal systems and decrypted some customer environment variables. Railway paused new builds and deployments while customers were told to rotate keys.
- OpenCode: an older system was accessed; API keys of 1.4% of users potentially exposed.
- LiteLLM: a malicious release pushed via three developers’ GitHub accounts could steal passwords and API keys; removed after about 2–3 hours.
Defensive tools mentioned: Z.ai’s GLM 5.3 reportedly found over 2,400 vulnerabilities, and its free Open Vuln Space scans a submitted repo and reports privately to maintainers (see the September landscape).
Open models you can run
| Release | What it is | Size / licence (as stated) |
|---|---|---|
| Drama 3 (Fish Audio) | TTS with emotion meta tags and cloning; #1 open TTS on Artificial Analysis | 35B MoE, 3B active; reported on 8 GB VRAM |
| Ming 3.0 TTS (Ant Group) | TTS with natural-language voice design, cloning, emotion and sound-effect tags, many languages and Chinese dialects | ~3 GB and ~18 GB (multi-speaker podcasts); MIT |
| Bark3 (Underdog) | Speech-to-text, free Hugging Face Space | 163 MB, Plus, 1.25 GB Pro; commercial use with attribution |
| MOSS Image (OpenMOSS) | 5B text-to-image, layout control, Base and SFT models | ~10 GB transformer; Apache 2.0 |
| Fibo 2.0 (Bria) | Natural-language image editor, multi-reference, restoration, colour correction | INT8 8.4 GB; non-commercial, gated access |
| FLUX 3 Image Klein (BFL) | 5.6B generate-and-edit, 4K, bounding boxes, ComfyUI support | ~10 GB; FLUX community licence (non-commercial; separate licence above about $5M revenue, per him) |
| WeVisOne-VL (Tencent) | Vision model: detection, counting, pointing, OCR, segmentation and alpha matting | 4B (19 GB) and 2B; Apache 2.0 |
| Hi3D 1.0 | Image or prompt to detailed 3D assets (v3 stays closed) | 7.74 GB |
| UniForge 3D | A few photos to a complete 3D model, filling unseen parts | Code and models released |
| InfiniSplat v2 | Multi-photo 3D scene reconstruction, including 1,000-frame drives | ~6.6 GB |
| PURE | 3D reconstruction from ordinary video | 13.2 GB, training scripts included |
| GaussiAnimate | Animating 3D characters with soft deformation (fur, clothing) | Code released |
| Matrix-Game 4.0 (Skywork) | Real-time playable game world: a game engine handles physics, the model renders; 83 fps at 480p on one consumer GPU; a coding agent edits game logic | Code and model released |
| Hy-Nomas (Tencent) | Text-controlled interactive world model, 720p 24 fps, minute-plus consistency | 34 GB base, 9 GB distilled INT8; Apache 2.0 |
| Multics | Multiplayer world model, trained on 4 players and scaling to 10 | ~16 GB |
| MosaiChunk, ATAS | Long-video memory for video models (one variant built on MiniMax H3) | Code released |
| StreamRender-H3 (Nvidia), Streaming Code World Model | Game-engine graphics restyled in real time by a video model | StreamRender over 200 GB, run on 8× GB200 |
| FastH3 Omni | 8-step MiniMax H3 reference-to-video, up to ~7× faster | ComfyUI, FP8, FP4 and Apple-silicon builds |
| Step 5 Preview (StepFun) | Agentic coding and reasoning LLM, 1M context, ~$1 per task via API | 600B MoE, 27B active, 650 GB |
| Nemotron 3.5 Super VL (Nvidia) | Fast multimodal model, 600+ tokens/s, training recipe released | 120B, 12B active, 75 GB |
| Ling-3.1-flash (inclusionAI) | Near GLM 5.3 / Kimi K3 / DeepSeek V4.1 on its benchmarks | 560B, 25B active, 574 GB |
| LFM2.5-24B-A2B (Liquid AI) | Small, fast agentic model, ahead of Qwen 3.6 and Gemma 4 in its class | 48 GB; mostly Apache 2.0 |
| Carbon-X (Hugging Face) | 14B DNA understanding and generation, from Qwen 3 | <30 GB; Apache 2.0 |
| Sage-A1.5 | 35B MoE scientific-research agent that writes cited reports with figures | 72 GB full, 22 GB Q4 GGUF |
| MemT 1.0 (MemTensor) | Persistent memory across conversations without re-reading history | Nano and 288B Flash variants |
| Ling-3.0-flash-Santé (inclusionAI) | Medical model, 124B / 5.1B active | 255 GB |
| RP1 (Roboparty) | Open-source humanoid robot, hardware designs included | GitHub |
The presenter calls Ling-3.0-flash-Santé an Anthropic release; the Hugging Face link he gives is inclusionAI’s.
Hosted releases
- Gemini 4 Argon (Google, public release week of Oct 18). Beats GPT-6 Astra and Opus 5.5 on several Google benchmarks; designed enzymes that worked in a lab; a quantum algorithm with a 100x cost reduction; video-to-robot skill transfer. 1M output tokens (“ten times” other frontier models). Third-party: #2 MazeBench, #3 Vending-Bench, #1 BlueprintBench (photo to floor plan), #1 Vals index and cheaper than #2 and #3, #3 KernelBench, near 98% on ARC-AGI-3 with a harness, #1 on clinical-image reasoning. Requires Pro or Ultra or a paid API plan. More in the October landscape.
- Muse Wonder and Muse Spark Open (Meta). Muse Wonder is fast and has a low hallucination rate on AA’s Omniscience test (it says “I don’t know” rather than inventing), but costs about 20M in 12 months, and personal details to download.
- Kimi K3.1 Pro (Moonshot). About 3T parameters, 1M context; on the AA index above Fable 5.1, GPT-6 and Gemini 4 at about $2 per task; strong Unreal, Blender and spreadsheet demos. On Kimi’s platform, paid plans and API now; open weights promised November 4. See Kimi K3.
- MiniMax M3.1 Flash. Fast, cheap agent model (the “Space Bunny” stealth model on OpenRouter), most cost-efficient on AutomationBench per MiniMax, close to the latest DeepSeek. API and MiniMax plans now; weights “later.”
- Kling 4. More dynamic motion, stereo audio, 10-bit HDR, 4K “in the near future,” 15 references, 10 keyframes. His view: close to, but not as good as, Seedance.
- Krea 3. 2K generation or edits in under 8 seconds, pixel-precise edits that leave the rest untouched, six references; paid, on Krea’s site; open-sourcing “briefly mentioned.”
- Ideogram Vector. Prompt or upload an image and get an editable vector (logos, illustrations, type); keeps much of a complex image’s detail. In the Ideogram app via “generate vector.”
- JST-2 (Viggle). Text or images to an explorable, consistent 3D world with animatable characters; beats Meshy and Tripo on its own win-rate test; free in the browser.
- Sky-001 (Skyfall). An “engineering world model”: breaks an object (a roadster, a jet engine) into 3D parts with CAD sketch steps, and generates floor plans from a building outline. Waitlist only.
- Dessa (Block). A speech-to-speech conversational model trained only on listening to conversations, with laughter, breathing and sarcasm; access limited to AI-safety researchers.
- Claude Duet (Anthropic). Proactive assistant on Opus 5.5 that summarises email, calendar and tasks each morning, monitors connected apps and runs tasks on its own virtual computer after you close the app; 6,000+ integrations including Gmail, Slack, Stripe and QuickBooks; US waitlist, Pro or Max. See Claude Duet.
- OpenAI research agents. Thousands of GPT agents analysed a mouse-brain gene-perturbation dataset in 32 minutes and cited 700+ papers against 11 for a single agent; Alzheimer’s and Parkinson’s leads are hypotheses only.
Research and other news (brief)
- OpenAI insider recording incident: an employee allegedly planted a $20 audio recorder behind a conference-room TV; arrested by the FBI on October 9, per the video.
- Google quantum: “magic state cultivation” on the Willow processor cut errors about 40x to 99.99% fidelity (paper in Science).
- Next-Concept Modeling (Stanford Hazy Research): text generated coarse-to-fine in parallel stages, up to 21x faster than competing diffusion language models on its tests.
- Robots: PingpongDuo (two humanoids playing doubles table tennis), Unitree H2X (a heavy-duty “guard” humanoid), RP1 (open hardware).
Update 2026-11-01 (Matt Wolfe, same week)
Source: Matt Wolfe, “AI News - 24 Things You Need To Know This Week” (raw/AI_News_-_24_Things_You_Need_To_Know_This_Week.md, https://www.youtube.com/watch?v=TVRCcMlvR10, fetched 2026-11-01). Sponsored in part by HyperAgent. Covers the same week as the October 25 episode; only items not already above are listed. “Voxtrot 3” in the captions is corrected to Voxtral 3.
- PHP exploitation numbers. Cloudflare CEO Matthew Prince: thousands of exploit attempts in the 10 hours before disclosure, tens of thousands within 45 minutes after, hundreds of thousands within two hours, “millions now.” Wolfe says OpenAI’s GPT-5.6 Sol found the bug and OpenAI shared a checker tool; Wiz called it “among the most severe vulnerabilities discovered in software history.” Fuller account in The Find-Fix Asymmetry and Black October.
- Codex “28 days” log, days 12-17. Day 12: more effective usage in Codex and Work via context reuse across sub-agents, longer caching and better routing. Day 14: edit an earlier message in a desktop-app thread; per OpenAI’s VB, all model output and tool calls after the edited message are discarded. Day 16: 12-hour prompt caching in the API, and for logged-in Codex and Work users cache is kept indefinitely, so resuming an old conversation hits the cache. Days 13, 15 and 17 were “resets.”
- OpenAI, smaller items. PIN-locked chats in ChatGPT (rolling out gradually); the Decisions API declared generally available, built on GPT-6 Luna for bounded answers (a probability, a choice from options, or a score). See Jev Decision Model for the competing product.
- Perplexity. Open-sourced its own decision models at 4B and 27B parameters; Perplexity Computer can now run fully on a Mac, using a Qwen 3.6 model with 36 GB or more of unified memory recommended.
- Sites blocking agents. After Amazon, Wolfe lists Yelp, Delta, JPMorgan Chase, Charles Schwab, cars.com and Amtrak as blocking AI agents; Shopify has declared its stores agent-friendly. Yelp separately launched calls to restaurants on your behalf to check wait times and menu items.
- Voice and video. Mistral’s Voxtral 3 voice agent diarises several speakers in real time, handles interruptions and calls tools. ElevenLabs’ Chimera combines Seedance 2.5 and ElevenLabs V4 to lip-sync an image to audio; in Wolfe’s test it ignored his framing (new shots, a stool, a mirror), lip sync was “decent”, and reference audio is capped at 30 seconds (the UI only tells you after generation).
- Images. Ideogram Vector: about 30 seconds for four editable SVG options, two of four usable in his test. Canva Craft is Canva’s first image model, built by the Leonardo AI team, in Canva’s Dream Lab tab and the Canva API (Wolfe discloses a small Canva investment). Rumour: “Aurora” model names in OpenAI’s API suggest a GPT Image 3.
- Other. Zoom launched enterprise deepfake-risk detection for calls; Microsoft published “Building Medical Super Intelligence” with diagnostic-accuracy claims against physicians (the model name was garbled in the captions); Unitree’s viral H2 “Kung Fu” video is, in Wolfe’s view, teleoperated and staged.
Try It
- If you deploy on Railway or route models through LiteLLM, rotate every API key and password stored in environment variables, and pin LiteLLM to a version released after the incident. ^[inferred] Check Railway’s and LiteLLM’s own advisories for the affected window.
- Voice-over for client video: test Ming 3.0 TTS (MIT) or Drama 3 locally against your paid TTS on one script, using their emotion tags. Check Drama 3’s licence first; the video does not state it.
- Product cut-outs and edits: WeVisOne-VL for alpha mattes (Apache 2.0); Fibo 2.0 only for non-commercial work; MOSS Image or Qwen Image 2.1 for commercial generation.
- Transcribe on a laptop: try Bark3’s free Space, then run the 163 MB model locally for meeting or podcast transcripts.
- Vector brand assets: run an existing raster logo through Ideogram Vector and check the paths in Illustrator or Figma.
- Cut repeat-context costs on OpenAI: if you resume long Codex or API sessions, the 12-hour API cache (Day 16 above) makes resuming within the window cheaper; structure prompts with a stable prefix to benefit.
- Long documents in one pass: use Gemini 4 Argon’s 1M output tokens for multi-file code or long-form drafts, on a paid plan.
Open Questions
- None of the security incidents are cross-checked against vendor advisories in this wiki; the 3%-of-all-critical-bugs chart has no named source.
- Licences not stated in the videos: Drama 3, Hi3D 1.0, Matrix-Game 4.0, Kimi K3.1 (until release), MiniMax M3.1 Flash.
- Kimi K3.1 Pro’s November 4 open-weight date is a promise.
- FLUX 3 Klein’s revenue threshold (“I think if you make over 5 million”) is the presenter’s recollection.
Related
- Frontier Model Landscape, October 2026 — Gemini 4 Argon, Muse Wonder and the open-weight wave in more depth
- Local Image Models in ComfyUI — MOSS Image and Qwen Image 2.1 — install tutorial for MOSS
- Local Video Models in ComfyUI — MiniMax H3 and LTX 2.5 — the H3 base that FastH3 Omni and StreamRender build on
- DeepSeek V4.1 Flash and DeepInfer — the DeepSeek release covered the same week
- Claude Duet — Anthropic’s proactive assistant
- Kimi K3 (Moonshot AI) — the predecessor to K3.1 Pro
- AI Voice — where the TTS releases fit